Agent Commerce Index (pilot)

In the 2 Nov 2026 pilot, AI models reached a checkout step on their own in 20 to 32 of 68 attempts each on 35 non-Shopify stores. In the 6 Nov 2026 snapshot, ARC's check got a checkout link back from 88.4% of Shopify stores (7,171 of 8,115) and 1.9% of stores on every other platform (49 of 2,583).

What it measures: whether an AI agent gets from a product page to a checkout step on a real store by itself, stopping before payment, and how many stores and protocols hand agents a checkout link at all.

Pilot, directionalv0.16 Nov 2026

Agent results: ARC-100 v1 pilot, 2 runs on 2 Nov 2026, 35 non-Shopify stores, 3 models called through an API. Store and protocol results: the 6 Nov 2026 open-data snapshot. How this is measured

A. Reached checkout, by model

Hand-verified checkouts out of every attempt, pooled over the pilot's 2 runs. The bar is the observed rate; the black line is the Wilson 95% interval. Every pair of intervals overlaps, so the pilot does not rank these models.

Reached checkout, by model (95% intervals)Reach-checkout rate per model in the ARC-100 v1 pilot, pooled over 2 runs on 35 non-Shopify stores. Claude Sonnet 5.5 32 of 68 (47%, 95% CI 36%–59%); Gemini 3.8 Flash 28 of 68 (41%, 95% CI 30%–53%); Qwen3-VL 235B 20 of 68 (29%, 95% CI 20%–41%).Claude Sonnet 5.5: 32 of 68 attempts reached checkout (47%), 95% interval 36% to 59%Claude Sonnet 5.547% · 32 of 68 · 95% CI 36%–59%Gemini 3.8 Flash: 28 of 68 attempts reached checkout (41%), 95% interval 30% to 53%Gemini 3.8 Flash41% · 28 of 68 · 95% CI 30%–53%Qwen3-VL 235B: 20 of 68 attempts reached checkout (29%), 95% interval 20% to 41%Qwen3-VL 235B29% · 20 of 68 · 95% CI 20%–41%0%25%50%75%100%
EmbedStatic SVG from the same data: <img src="https://www.arcreport.ai/embed/index-a.svg" alt="Reached checkout, by model">
Data table
ModelReached checkoutAttemptsRate95% interval
Claude Sonnet 5.5326847%36%–59%
Gemini 3.8 Flash286841%30%–53%
Qwen3-VL 235B206829%20%–41%

B. Success vs. cost per checkout

Each point is one model: pooled success rate against model token cost per checkout (the total cost of both runs divided by pooled checkouts). Vertical lines are 95% intervals. Up and to the left is better.

Success vs. cost per checkoutPooled reach-checkout rate against model cost per checkout (total model cost over both runs divided by pooled checkouts). Claude Sonnet 5.5: 47% at $0.17; Gemini 3.8 Flash: 41% at $0.07; Qwen3-VL 235B: 29% at $0.03. Vertical lines are 95% intervals.0%10%20%30%40%50%60%$0.00$0.05$0.10$0.15$0.20Model cost per checkout (pooled)Claude Sonnet 5.5: 47% reached checkout (32 of 68), $0.17 model cost per checkout, 95% interval 36% to 59%Claude Sonnet 5.5Gemini 3.8 Flash: 41% reached checkout (28 of 68), $0.07 model cost per checkout, 95% interval 30% to 53%Gemini 3.8 FlashQwen3-VL 235B: 29% reached checkout (20 of 68), $0.03 model cost per checkout, 95% interval 20% to 41%Qwen3-VL 235B

Cost per checkout here is pooled over both runs, so it differs slightly from the per-run figures on the pilot page.

EmbedStatic SVG from the same data: <img src="https://www.arcreport.ai/embed/index-b.svg" alt="Success vs. cost per checkout">
Data table
ModelRateModel cost, both runsCost per checkout
Claude Sonnet 5.547%$5.29$0.17
Gemini 3.8 Flash41%$2.06$0.07
Qwen3-VL 235B29%$0.60$0.03

C. Where agents stop

Every pilot attempt by the furthest stage it reached. The pilot records four stages per attempt; finer steps (product not found, wrong variant, sign-in wall) are not recorded yet. Counts are printed on each segment, so colour is not the only signal.

Where agents stopPooled outcome of every attempt per model, furthest stage first. Claude Sonnet 5.5: reached checkout 32, cart only, no checkout step 12, did not get the product into the cart 18, blocked at the door 6 (n = 68); Gemini 3.8 Flash: reached checkout 28, cart only, no checkout step 15, did not get the product into the cart 19, blocked at the door 6 (n = 68); Qwen3-VL 235B: reached checkout 20, cart only, no checkout step 20, did not get the product into the cart 22, blocked at the door 6 (n = 68).Reached checkoutCart only, no checkout stepDid not get the product into the cartBlocked at the doorClaude Sonnet 5.5n = 68Claude Sonnet 5.5: Reached checkout, 32 of 68 (47%)32Claude Sonnet 5.5: Cart only, no checkout step, 12 of 68 (18%)12Claude Sonnet 5.5: Did not get the product into the cart, 18 of 68 (26%)18Claude Sonnet 5.5: Blocked at the door, 6 of 68 (9%)6Gemini 3.8 Flashn = 68Gemini 3.8 Flash: Reached checkout, 28 of 68 (41%)28Gemini 3.8 Flash: Cart only, no checkout step, 15 of 68 (22%)15Gemini 3.8 Flash: Did not get the product into the cart, 19 of 68 (28%)19Gemini 3.8 Flash: Blocked at the door, 6 of 68 (9%)6Qwen3-VL 235Bn = 68Qwen3-VL 235B: Reached checkout, 20 of 68 (29%)20Qwen3-VL 235B: Cart only, no checkout step, 20 of 68 (29%)20Qwen3-VL 235B: Did not get the product into the cart, 22 of 68 (32%)22Qwen3-VL 235B: Blocked at the door, 6 of 68 (9%)60%50%100%
EmbedStatic SVG from the same data: <img src="https://www.arcreport.ai/embed/index-c.svg" alt="Where agents stop">
Data table
ModelReached checkoutCart only, no checkout stepDid not get the product into the cartBlocked at the doorAttempts
Claude Sonnet 5.5321218668
Gemini 3.8 Flash281519668
Qwen3-VL 235B202022668

D. The platform gap

Stores in the 6 Nov 2026 snapshot where ARC's automated HTTP check got a checkout link back (it stops before payment), out of stores whose check finished. Shopify: 88.4% (7,171 of 8,115). Every other platform: 1.9% (49 of 2,583).

The platform gap: checkout link returned, by platformStores in the 6 Nov 2026 snapshot where ARC's automated HTTP check got a checkout link back, by platform. Shopify: 7,171 of 8,115 (88.4%); Every other platform: 49 of 2,583 (1.9%); Not detected: 21 of 1,263 (1.7%); Platform not recorded: 28 of 831 (3.4%); Salesforce Commerce Cloud: 0 of 168 (0%); WooCommerce: 0 of 107 (0%); Adobe Commerce / Magento: 0 of 78 (0%); BigCommerce: 0 of 52 (0%); Squarespace: 0 of 41 (0%); Wix: 0 of 24 (0%); Other named platforms (Ecwid, Lightspeed, PrestaShop): 0 of 19 (0%).Shopify: ARC's check got a checkout link back from 7,171 of 8,115 stores (88.4%)Shopify88.4% · 7,171 of 8,115Every other platform: ARC's check got a checkout link back from 49 of 2,583 stores (1.9%)Every other platform1.9% · 49 of 2,583Every other platform, one by oneNot detected: ARC's check got a checkout link back from 21 of 1,263 stores (1.7%)Not detected1.7% · 21 of 1,263Platform not recorded: ARC's check got a checkout link back from 28 of 831 stores (3.4%)Platform not recorded3.4% · 28 of 831Salesforce Commerce Cloud: ARC's check got a checkout link back from 0 of 168 stores (0%)Salesforce Commerce Cloud0% · 0 of 168WooCommerce: ARC's check got a checkout link back from 0 of 107 stores (0%)WooCommerce0% · 0 of 107Adobe Commerce / Magento: ARC's check got a checkout link back from 0 of 78 stores (0%)Adobe Commerce / Magento0% · 0 of 78BigCommerce: ARC's check got a checkout link back from 0 of 52 stores (0%)BigCommerce0% · 0 of 52Squarespace: ARC's check got a checkout link back from 0 of 41 stores (0%)Squarespace0% · 0 of 41Wix: ARC's check got a checkout link back from 0 of 24 stores (0%)Wix0% · 0 of 24Other named platforms (Ecwid, Lightspeed, PrestaShop): ARC's check got a checkout link back from 0 of 19 stores (0%)Other named platforms (Ecwid, Lightspeed, PrestaShop)0% · 0 of 19

Platforms with fewer than 20 finished checks are grouped. A 0% bar is a measured zero: the check finished and no link came back. Counts from manifest.json and store_answers.csv; see which figure is which.

EmbedStatic SVG from the same data: <img src="https://www.arcreport.ai/embed/index-d.svg" alt="The platform gap">
Data table
PlatformCheckout link returnedStores checkedShare
Shopify7,1718,11588.4%
Every other platform492,5831.9%
Not detected211,2631.7%
Platform not recorded288313.4%
Salesforce Commerce Cloud01680%
WooCommerce01070%
Adobe Commerce / Magento0780%
BigCommerce0520%
Squarespace0410%
Wix0240%
Other named platforms (Ecwid, Lightspeed, PrestaShop)0190%

E. The protocol funnel

For UCP and MCP, how many checked stores pass each stage. A hatched cell is a stage ARC does not measure for that protocol: it is not a zero. UCP's checkout-link stage is the same check as the platform gap above.

Protocol funnel: declared, valid, usable, checkout linkProtocol funnel from the 6 Nov 2026 snapshot. UCP: declared not measured, valid 8,193 of 10,698, usable 7,635 of 10,698, checkout link 7,220 of 10,698; MCP: declared 7,751 of 10,698, valid 7,751 of 10,698, usable 7,751 of 10,698, checkout link not measured.DeclaredValidUsableCheckoutlinkUCPUCP, Declared: not measured. /.well-known/ucp answered HTTP 200 with a bodyNotmeasuredUCP, Valid: 8,193 of 10,698 stores (76.6%). Measured. Profile passed ARC's validation8,193of 10,69876.6%UCP, Usable: 7,635 of 10,698 stores (71.4%). Measured. The UCP catalog tool returned the tested product7,635of 10,69871.4%UCP, Checkout link: 7,220 of 10,698 stores (67.5%). Measured. UCP create_cart returned a checkout link7,220of 10,69867.5%MCPMCP, Declared: 7,751 of 10,698 stores (72.5%). Measured. /api/mcp answered a tools/list request with tools7,751of 10,69872.5%MCP, Valid: 7,751 of 10,698 stores (72.5%). Measured. tools/list parsed with at least one named tool7,751of 10,69872.5%MCP, Usable: 7,751 of 10,698 stores (72.5%). Measured. A tool or catalog answered on /api/mcp (on Shopify today, the policy and FAQ tool; or a catalog tool returned the tested product)7,751of 10,69872.5%MCP, Checkout link: not measured. Of stores whose /api/mcp lists a cart tool, the cart tool returned a checkout link (stopped before payment). Not measured while no checked store lists one there.Notmeasured

From protocol_matrix.csv. Every protocol and stage, with definitions, is on the Protocol Matrix.

EmbedStatic SVG from the same data: <img src="https://www.arcreport.ai/embed/index-e.svg" alt="The protocol funnel">
Data table
ProtocolDeclaredValidUsableCheckout link
UCPNot measured8,193 of 10,698 (76.6%)7,635 of 10,698 (71.4%)7,220 of 10,698 (67.5%)
MCP7,751 of 10,698 (72.5%)7,751 of 10,698 (72.5%)7,751 of 10,698 (72.5%)Not measured

Methodology and independence

  • Pilot, directional. Agent results come from 2 runs on 35 non-Shopify stores of ARC-100 v1; 64 Shopify stores were not run. Models were called through an API in ARC's browser harness, not through any consumer app.
  • Stops before payment. No step pays or places an order. "Reached checkout" means a checkout step loaded with the right product.
  • Two sources, two questions. Charts A to C are agent attempts in the pilot. Charts D and E are ARC's automated HTTP check across the snapshot. They count different things and are never added together.

Pilot results and method · Benchmark methodology · How ARC checks a store · Metric definitions · Method changelog · Open dataset

  • ARC takes no payment from any agent, platform or protocol company.
  • ARC is not affiliated with any AI company.

Cite this (CC-BY 4.0)

Free to reuse with attribution: ARC, arcreport.ai.

ARC (2026). ARC Agent Commerce Index (pilot) v0.1. Data as of 6 Nov 2026. https://www.arcreport.ai/agent-commerce-index

@misc{arc_agent_commerce_index_20261106,
  author = {{ARC}},
  title  = {ARC Agent Commerce Index (pilot) v0.1},
  year   = {2026},
  note   = {Data as of 6 Nov 2026. Licence CC-BY-4.0},
  url    = {https://www.arcreport.ai/agent-commerce-index}
}